You ask ChatGPT for keywords for a new product, and it returns a tidy list with search volumes and difficulty scores. Then you try to find where those numbers came from, and nothing is there. That is the real problem with how to do SEO keyword research using ChatGPT Work: the output looks finished long before it is true. Work can write a clean brief, but it cannot see live volume, difficulty, or a results page unless a plugin supplies those cells. Search Engine Land said years ago that ChatGPT is not a keyword tool, and that essay still ranks. Demand for the topic is real. US searchers look up "chatgpt work" about 33,100 times a month, while "chatgpt keyword research" sits at 20 (DataForSEO Labs, 31 Oct 2026). Queries that pair Work with keyword research have no measured volume yet. The old warning still holds, with one update: it is true until Metaflow is connected, because a connected plugin gives Work real cells to quote.
Uploading files does not close the gap. Attach a month-old Ahrefs CSV and ask Work to "pick winners," and every blank cell is still a guess the model will quietly fill. The honest way to do SEO keyword research using ChatGPT Work has three moves: start a research job through the plugin, quote only attributed figures, and stop at UNKNOWN when a figure has no source. This guide walks through each move and shows the board and brief you should end up with. It also explains how to put the refresh on a schedule so next month's run builds on this month's work.
TL;DR
- How to do SEO keyword research using ChatGPT Work starts with a plugin, not a prompt pack.
- Chat seeds wording. Work finishes one decision brief.
- Metaflow runs the research. Work writes the board.
- Label missing metrics UNKNOWN. Do not let the model fill them.
- Refresh on a schedule. Do not rebuild the prompt each month.
What how to do SEO keyword research using ChatGPT Work really requires
The workflow needs three inputs before you open a thread. If one is missing, the output will look complete and still be unreliable.
- A business frame: who you help, the conversion that matters, and the market and language you are targeting.
- An inventory: the URLs you already have, with owners and last-updated dates if the team has recorded them.
- A plugin that can pull live keyword and SERP facts. SERP means search engine results page, the live list of pages that currently rank for a query.
The first two inputs tell Work what you are trying to win and what you already own. The third supplies the evidence. Without it, how to do SEO keyword research using ChatGPT Work collapses into the 2023 method: brainstorm now, validate later in a real tool. That method burns your Work usage pool on thinking Chat already does for free.
SEO plugins for ChatGPT Work compares plugins that only look things up with plugins that run a full research engine. This page assumes you have picked the engine. Ahrefs in the same thread is optional when you want a second table to cross-check. It is not a substitute for a job that ends in a page route, meaning a decision to refresh, create, consolidate, or skip a page.
Files, Chat, and a plugin
Each path below suits a different stage, and the table shows where each one stops being trustworthy. Read the "Stop when" column as the signal to switch paths, not as a criticism of the path.
| Path | Use it when | Stop when |
|---|---|---|
| Chat | You need ten seed phrasings | You start assigning volume |
| Work, files only | You already trust every cell in the export | A cell is blank |
| Work plus Metaflow | You need live research and a brief | You are tempted to publish from the thread |
The third row is the workflow in this tutorial. The first two rows remain useful, since Chat is a good way to generate wording and a clean export is a good way to check a plugin. The important point is that every row has a stopping rule, and the rule for files-only work is the blank cell.
The next table sets the standard for what counts as evidence in the brief. It splits each type of cell into what you can call proven and what must be labeled UNKNOWN.
| Cell | Proven | UNKNOWN |
|---|---|---|
| Volume / difficulty | Plugin returned it with a date | Model filled a blank |
| SERP page type | You or the plugin saw the live type | "Probably a guide" |
| Inventory | URL exists in the list you attached | "We should have a page" |
| Claim | Source in the brief | A competitor boast |
The practical lesson is that "proven" always has a source and a date, and "UNKNOWN" is any cell where the model supplied a guess. A confident phrase like "probably a guide" belongs in the UNKNOWN column.
If more than a third of the board is UNKNOWN, you do not have a calendar. You have a research debt list. Assign an owner to each gap, then run the research again once the cells exist. Do not brief writers on a debt list and call it a plan.
A second pass is allowed. A second invention is not. When a stakeholder asks why a row is empty, show them the table. That conversation is part of the job, and the plugin cannot have it for you.
Keep the first board to one product and one market. A global dump in a single Work thread hides the one cluster you could actually ship this month. How to do SEO keyword research using ChatGPT Work is a decision tool, and a good decision often means saying no to most of the list.
OpenAI's Work help says Work is meant for research and finished deliverables. How to do SEO keyword research using ChatGPT Work is both, provided the research is real.
How to do SEO keyword research using ChatGPT Work in one pass
A single pass means one goal, one research job, and one board at the end. Switch to Work, install Metaflow, and connect the workspace. Open a project if the brand brief and URL inventory already live there, so the job can read them. Then send one goal rather than a strategy offsite, because a narrow goal produces a board you can actually check.
The start_work prompt
The prompt below is what tells the plugin to start a research job. It does three things: it defines the unit of work (a cluster, which is a group of queries sharing one searcher job), it lists the fields you want back, and it forbids invented numbers.
> How to do SEO keyword research using ChatGPT Work for [product] in [country/language]. Cluster by searcher job, not wording. For each cluster: primary query, support queries, intent, volume, difficulty, live SERP page type, refresh or new URL, owner if known, UNKNOWN if unproven. Do not invent metrics. Attribute every number to Metaflow.
Once the plugin starts the job, it reports the steps it is taking. Poll the step it quotes, answer any questions it asks, and read what it says it is about to do. Do not approve anything that looks like a publish action or a crawl of a property you do not own.
When findings return, Work should write a candidate board, not a blog post. A candidate board is a table with one row per searcher job. If two intents hide under one wording, split the row. If the live page type is a tool and you planned an essay, mark the mismatch. That mismatch is the whole point of how to do SEO keyword research using ChatGPT Work against a live SERP instead of a guessed list.
Read the board before you ask for prose. A finished first pass has a market, a product, and a row for each searcher job you would actually staff. Volume and difficulty carry a source. The live page type is what ranks now, not what you wish ranked. Refresh, new, consolidate, or no page is filled in for every row. Anything the plugin could not prove says UNKNOWN, with a person's name next to it.
If a row is a pile of synonyms sharing one intent, split it or delete it. This workflow fails in a familiar way at this point. The thread looks complete because it is long, but the board is still a dump because two jobs are hiding under one wording.
Programmatic SERP analysis is the deeper crawl path when the board is too large for one Work thread. Use it after the first pass, not instead of it, since you need the first board to know which clusters deserve a deeper look.
Turn findings into a Work deliverable
A board tells you what exists. A deliverable tells a writer what to do. Ask Work for one artifact, a decision brief, and require these fields:
- Primary query and support queries
- Searcher job in one sentence
- Evidence cells with sources
- Page route: refresh, new, consolidate, or no page
- Writer notes: what must be proven, what is out of scope
- Open questions with owners
This list turns how to do SEO keyword research using ChatGPT Work from a dump into a decision. We call the structure the Candidate Board framework: one row per searcher job, evidence behind every figure, a page route, and an owner for every UNKNOWN. A keyword dump is what you get when the thread stops at a table. A decision brief is what you get when Work has to commit to refresh, new, consolidate, or no page.
For background on what volume and difficulty scores actually measure, Ahrefs publishes its own keyword research guide. Read it before you argue about a number, because tools define difficulty differently and the brief should name which definition it used.
Three failures show up in the first real run. Each has a cause you can trace and a change that prevents it.
- The model fills a blank volume cell because the plugin timed out and nobody labeled the gap. The brief then carries a number with no source. The fix is a rule that any timed-out lookup returns UNKNOWN, plus a re-run request.
- Two intents share one wording, and the brief still proposes one URL. This surfaces when the writer's draft satisfies one audience and bounces the other. The fix is to split the row before the brief is written.
- The live page is a tool and the plan is an essay, and the mismatch never gets a row of its own. It surfaces months later as a page that never ranks. The fix is to record page type in the route column.
Treat each of those as a stop, not a polish pass. How to do SEO keyword research using ChatGPT Work is finished when a second person can defend the route without reopening the chat. If they cannot, the board is not done.
The complete AEO guide is the longer treatment when the board also has to answer how AI search cites the URL. The Dots guide for marketing covers the standing job you will want after you like the first board. Scheduling the first design pass is how UNKNOWN columns come back next month with new names and the same missing sources, so schedule only after the design is settled.
A stakeholder who asks for the best keyword should see the board and the open questions, not a single phrase with a made-up difficulty score. The information gain is the refusal: no source, no cell, no page. That rule separates this workflow from the 2023 prompt packs that still rank for the parent query. Those packs produce wording. How to do SEO keyword research using ChatGPT Work produces a route you can hand to a writer with the evidence attached.
When the inventory and the live page type disagree, the brief should say so in the route column instead of smoothing it into a content calendar. Doing this well earns the handoff only after that disagreement is visible. The aim is one URL and one owner, then the next Work chat.
When how to do SEO keyword research using ChatGPT Work should stop
Stop before draft copy for the public page, and stop before anyone "just publishes the outline." The research job ends at the brief. Content-led growth is the system that takes a brief to a URL, so it picks up where this work ends. If your next job is ads or outbound rather than another cluster, marketing plugins for ChatGPT Work is the hub to start from.
If a stakeholder wants "the best keyword," show them the board and the UNKNOWN column. That discipline is where the information gain lives, and it is the reason the brief can be trusted.
Schedule the refresh without rebuilding the prompt
Keyword data goes stale, so a one-time board is a snapshot, not a system. Save the goal as a scheduled Work task, or hand it to a Dot for SEO once you like the first board. A Dot is a standing job that wakes on its own. Installing the plugin does not create the schedule, so you still write the trigger and the timezone.
A monthly refresh should only reopen clusters where volume, SERP type, or inventory changed. It should not invent a new strategy from a blank Chat window.
Write the trigger the way you would brief a colleague who was not in the original thread. Name the product, the market, the file that holds last month's board, and the rule for what counts as a change. A volume swing with no change in the live page type is a note, not a new URL. A new page type on a query you already rank for is a route change, and it belongs in the brief before anyone opens a doc. A cluster that was UNKNOWN because the plugin lacked the market should stay UNKNOWN until that market is connected.
The refresh is allowed to be shorter than the first board. If it comes back the same length every month, with new wording and the same missing sources, the schedule is rerunning the prompt instead of inheriting the work.
How to do SEO keyword research using ChatGPT Work stays useful after the first week only if that inheritance is real. Store the board where the next run can read it. Put the timezone on the task so Monday in the client's market is the Monday you meant. Then read the diff, not the whole universe. The person who owns the UNKNOWN cells should see their names again, not a fresh list of questions the model invented because the prior answers were trapped in Chat.
This approach feels new because Chat trained teams to treat research as wording, and Work made research a deliverable. The fear that follows is a calendar built on invented difficulty. The cure is to encode the job into a workflow so the next cycle inherits the board.
The tension most teams feel by the third month is accountability. Someone asks why a row says UNKNOWN or why a page was routed to "new," and nobody can find the chat where that was decided. Debugging a bad calendar then means re-reading old threads, and every campaign starts by resetting what the last one learned.
That pain goes away when operator judgment is written down once as a skill, a workflow, or an agent, with stable context behind it. The rules you applied by hand, such as no source meaning no cell and mismatched page types getting their own row, then run the same way each cycle, and the work compounds instead of vanishing with the chat window. Discovery and execution also belong in one place, so a good exploration becomes a durable growth system rather than a one-off.
Metaflow is the natural place for that handoff. You explore freely in Chat, then solidify what worked into skills that carry your research rules, agents that run the standing job, and a flow that connects the board to the brief. Next month's refresh then starts from the last board instead of a blank page.
Frequently Asked Questions
Can ChatGPT Work do keyword research?
Yes, but only the thinking half of how to do SEO keyword research using ChatGPT Work. Work can cluster queries by searcher job, compare them to your URL inventory, and write a decision brief, yet it cannot see live volume, difficulty, or a results page by itself. A plugin has to supply those figures, and Metaflow is one way to connect that research to Work. Without a plugin you are back to the 2023 method of generating ideas now and validating them later in a separate tool.
Does ChatGPT Work invent search volume?
It can, and it will if you leave blanks for it to fill. When you ask for volume without a data source, or attach a CSV with empty cells, the model tends to produce a plausible number rather than admit the gap. The defense is a standing rule that every unproven figure is labeled UNKNOWN and that only attributed numbers enter the brief. A guardrail like this can live in a Metaflow skill, so it applies on every run without being retyped.
Do I need Ahrefs for keyword research in Work?
No, you only need it if you want a second lookup table in the same thread. The research job itself needs a plugin that returns live keyword and SERP data with a date, and that source can be Metaflow or another engine you trust. Many SEO leads keep Ahrefs for ad hoc checks and for cross-checking a surprising number. Skipping every data source and still calling how to do SEO keyword research using ChatGPT Work finished is the one choice that does not hold up.
Can I use Search Console with ChatGPT Work?
Yes, by exporting the data, because there is no native Search Console plugin that replaces an export. Attach a Search Console Performance export, or land the data where your research workflow can read it, for example in Metaflow. First-party clicks and impressions beat third-party volume for queries you already rank on, since they reflect your own site rather than a modeled estimate. Use both sources when you have both, and note which one each cell came from.
Should I schedule keyword research in ChatGPT Work?
Yes, but only after the first board is one you would defend. Schedule the refresh, not the first design pass, since scheduling a bad design just repeats its gaps every month. A Dot suits the job when it should wake on its own, and Work suits it when you still want to steer the artifact. Whichever you choose, store last month's board where the next run can read it so only the changes get reopened.
